A novel federated multi-view clustering method for unaligned and incomplete data fusion

Abstract

Recently, federated multi-view clustering (FedMVC) has emerged as a powerful tool to uncover complementary cluster structures across distributed clients, gaining significant attention in the realm of data fusion. While FedMVC methods have adeptly addressed the challenges of feature heterogeneity among various clients, achieving notable success in controlled environments. Their applicability often hinges on the assumptions of strict alignment and data completeness across multi-view clients. These assumptions, unfortunately, are not always consistent with real-world conditions. Specifically, practical applications often come with (1) unaligned multi-view data and (2) missing data. Current FedMVC methods struggle to effectively address these challenges. To bridge this gap, this paper presents FCUIF, a novel method that eliminates the need for data alignment and completeness assumptions. To tackle unaligned data, FCUIF leverages both sample commonality and view versatility to adaptively generate alignment matrices, ensuring effective cross-view alignment. For the challenge of missing data, FCUIF uses an unsupervised technique to evaluate and refine imputation quality, efficiently handling various scenarios of incomplete multi-view data. Our extensive experiments using four public datasets demonstrate FCUIF’s superior performance when dealing with unaligned and incomplete multi-view data. The code is available at https://github.com/5Martina5/FCUIF.